Skills

Skills are instructions an agent loads when the work calls for them. Agents can find and add published skills with your approval, write their own, and every change is kept as a version.

Updated 22 October 2026

A skill is a set of instructions an agent follows for one kind of work: a house style for release notes, a design taste, the steps for a tool your team uses. An agent sees a one-line description of every skill it carries and loads the full instructions when a request matches.

Every agent comes with Bolter's own skills. On top of those, a skill can be added to one agent, or to the whole workspace so every agent carries it.

Where skills come from

• Published skills. Most published skills live on GitHub, and the skills.sh index lists them. Any address npx skills add takes works in Bolter too: owner/repo, owner/repo@skill-name, owner/repo/path/to/skill, or a github.com or skills.sh link.
• Skills an agent writes. Teach an agent a way of working and ask it to keep it ("save this as a skill"). It writes the skill and uses it from then on. Ask it to change the skill later and it saves a new version.
• Skills an agent learns. When an agent has to work out how to do something by trial and error, such as an app, connector or API that only works when called a certain way, it saves the steps that worked as a skill once the task is done and tells you in a line. For something the whole workspace uses it saves a workspace skill, so other agents get it right first time. It may also add a short page to workspace memory.
• Skills you write. In a Skills tab, choose Write a skill, give it a name, a line saying when to use it, and the instructions.

Asking an agent to find one

Ask in plain words: "find a skill that would give you better frontend design taste and add it". The agent searches the published skills, picks one and asks to add it. You can also name one directly: "add Leonxlnx/taste-skill".

A skill is instructions the agent will follow, so a published skill is never added without a person approving it. The agent posts a card in the chat. Choose Review to read the whole skill, see where it comes from (at the exact commit), and see what the safety check found. Then:

• Approve adds it.
• Approval scope defaults to this version only. Choose Always allow for a GitHub owner or repository, then Approve & always allow, to trust that source from now on: later skills and updates from it are added without asking, as long as the safety check finds nothing wrong. Only workspace owners and admins can always allow a source.
• Decline adds nothing.

Until it is approved the agent cannot use the skill, and it says so.

Who decides

• A skill for one agent: anyone who can manage that agent (its creator, and workspace owners and admins).
• A workspace skill: workspace owners and admins.

When an agent asks, or an update from a source waits, everyone who can decide also gets a notification, so a request made in a chat they are not in still reaches them. A skill a person adds in Skills settings is the exception: if the safety check holds it, it waits in that Skills tab with what the check found, and no one is notified.

The safety check

Before any skill reaches an agent, a model reads it and points out anything clearly problematic: telling the agent to send secrets or private data somewhere, to ignore its rules or approvals, hidden instructions, or running code downloaded at run time. It quotes the exact lines so you can judge for yourself. If the check could not run, the review says so and never treats the skill as checked. A skill an agent writes for itself takes effect straight away when the check finds nothing; if the check flags something, a person approves it first.

skills.sh also publishes audits from its security partners for many skills; when there are some, the review shows them.

Updates

A skill added from GitHub stays connected to its source. Bolter checks for a new version a few times a day:

• From an always-allowed source, a new version the check finds clean is used straight away.
• Otherwise it waits for review, showing exactly what changed, and the agent keeps using the version you approved until someone decides.

Turn Keep up to date off on a skill to stop following its source. Editing a published skill makes it your own copy, which no longer updates.

Versions

Every change to a skill is a new version, whoever made it. Open a skill in a Skills tab to see its versions, what changed between them, who made or approved each one, and to Restore an earlier one. Restoring a published skill also stops it updating, so the version you chose stays in use.

Where to find them

• One agent: open the agent's profile and choose Skills. You see its own skills, the workspace skills it carries, and anything waiting for review.
• The workspace: open the workspace and choose Skills. You see the workspace skills and the always-allowed sources, and admins can remove a source there.

Turn a skill off to stop agents using it without removing it. Removing a skill keeps its history.

When an agent's own skill and a workspace skill share a name, the agent uses its own.

Skills that run scripts

Some skills carry scripts and data files as well as instructions. When an agent writes code on a computer, its skills are there too, so a coding agent can read and run them. In a chat, an agent that needs a skill's scripts copies the approved version onto one of its computers first and runs it from there. It never downloads the skill again from where it came from, because that could run a version nobody approved.

When an agent loads a skill, the chat shows it as a step with the skill's name.